Linear algebra, calculus, probability and statistics, taught as the tools you reach for when a model will not train. Two modules here are ours rather than the source courses: array programming, which every lab assumes, and numerical stability, which nothing teaches and everything depends on.
Arrays, shapes, broadcasting and the loop you should not write. Every lab in this curriculum is written against this.
Vectors, systems, matrices, transformations and eigenvectors, each one arriving before the thing that needs it.
Derivatives, integrals, gradients and the descent that uses them, ending at Newton's method.
Events, Bayes, random variables, distributions and the covariance matrix that PCA and anomaly detection both need.
Sampling, estimation, confidence, hypothesis testing and the A/B test they add up to.
Floating point, overflow, the log-sum-exp trick and the reason your loss became NaN.
Supervised learning end to end, then the diagnostics that tell you what to fix, then trees, clustering, recommenders and reinforcement learning. Gradient boosting still wins more production problems than deep learning does.
The supervised setup, the cost function, and gradient descent doing the work.
Decision boundaries, the logistic link, and a cost function that comes from the likelihood.
What overfitting is, what regularization does about it, and how to choose without fooling yourself.
Neurons, layers, forward propagation and a training loop, before the deep learning track takes it apart.
The diagnostics that tell you what to fix, and the loop that turns them into decisions.
Purity, splits, forests and boosting -- the family that still wins on tabular data.
Clustering without labels, and finding the rare event you have almost no examples of.
Collaborative filtering, content-based retrieval and the two-stage system they become.
Returns, policies, the Bellman equation and the deep Q-network that approximates it.
Networks from scratch, everything that makes a real one converge, and the two architectures that carried the field: convolutions for vision and attention for sequences. Ends at the transformer, implemented from the paper.
The forward and backward pass built from nothing, then stacked until it is deep.
Splits, the bias-variance recipe, regularization, initialization and the gradient check that finds the bug.
Mini-batches, momentum, Adam, schedules, and the hyperparameter search that finds the run.
One number to optimize, a human yardstick to beat, and an error analysis that says what to do next.
The convolution, the layer built from it, and the architectures that made depth work.
Finding where the object is, drawing its outline, and telling one face from another.
Recurrence, backpropagation through time, and the gates that fixed its memory.
Turning words into vectors whose geometry means something -- and the tokenizer underneath.
The bottleneck, the attention that removed it, and the architecture that dropped recurrence entirely.